Measured report · 03

Can a citation-first RAG service clear one million requests at low latency?

We exercised a Go retrieval API through a public GCP load balancer into Cloud Run and Cloud SQL, retained every failed stage, and priced the measured allocation separately from the production baseline.

1,000,000

successful responses

976.8 RPS

sustained throughput

13.45 ms

end-to-end p95

HTTP errors

The final mixed run completed one million attempts in 17.1 minutes with 94.97% cache hits. Five percent of traffic deliberately bypassed the cache, so Cloud SQL remained in the measured path.

Cloud Monitoring counted the same one million requests. The service peaked at 5 active instances while Cloud SQL averaged 11.2% CPU utilization.

Request path

A narrow production wedge for field service

Stable manufacturer fault codes take an exact-match lane; broader questions fall through to PostgreSQL full-text and pgvector retrieval.

Architecture diagram from client through a global load balancer and Cloud Run to an in-process cache or Cloud SQL grounded retrieval — Measured in us-east1 with a global external Application Load Balancer, serverless NEG, Cloud Run and PostgreSQL 16. The temporary database was zonal; the published production cost model uses regional HA.

Gates before scale

Fast retrieval was rejected until it found the right source

The first ranker was fast but recovered the expected synthetic source in the top three only 33.8% of the time.

FAILED · HYBRID RRF V1

33.8%

recall@3 across 500 deterministic queries

PASSED · EXACT CODE + HYBRID

100%

recall@1 and recall@3 on the same query construction

The corpus is deterministic and synthetic: 20,000 troubleshooting chunks with unique source IDs and fault codes. The result proves source recovery for this test construction, not maintenance safety, semantic answer quality or field accuracy.

Tail latency

The failed 1,000-RPS burst stays in the chart

A first burst scaled into cold per-instance caches and reached 221.51 ms p95. The immediate warm repeat and final mixed run passed.

Horizontal bar chart of p95 latency for failed scale-out, warm cache, forced database, 100,000-request and one-million-request stages — The forced-cache-miss run completed 10,000 database-path requests at 24.96 ms p95. The one-million mixed run measured 18.07 ms p99 and 39.70 ms p99.9.

Cost boundary

$0.85 marginal upper bound per million

Measured Cloud Run allocation, request charges, response-body load-balancer processing and North America internet egress. Free tier and discounts are excluded; request-header bytes are not captured.

$239.46

1M requests / month

$239.46 per million

$242.86

5M requests / month

$48.57 per million

$281.17

50M requests / month

$5.62 per million

Production model

The database baseline matters more than request volume

A regional-HA Cloud SQL database, one global forwarding rule and one warm Cloud Run instance total $238.61 per month before traffic.

Bar charts of estimated monthly RAG retrieval platform cost and cost per million requests at one, five and fifty million requests

At five million retrieval questions per month the modeled platform total is $242.86. That figure excludes LLM inference, ingestion, observability retention, support labour, taxes and any enterprise resilience work beyond the stated baseline.

Cloud Run pricing Cloud SQL pricing Load Balancing pricing

Download evidence

Raw results, method and reproduction commands

No email gate. The failed ranker, edge propagation, cache warmup and client-timeout runs are retained beside the passing results.

All raw runs and pricing inputs (JSON)Frozen result summary (JSON)Load-test results (CSV)Methodology and limitations Reproduction commands

Have repetitive, citation-sensitive service questions? Start by measuring retrieval quality on your actual documents and escalation rules.

Plan a production RAG gate